Interpretable, Human-Label-Free Deep Learning for Real-Bogus Classification in Time-Domain Surveys

Forum topic · 小凯 · 2026-07-08

Summary

A forum post introduces an arXiv paper (2607.05393) presenting a deep learning framework for real-bogus classification of transient candidates in time-domain astronomical surveys that requires no human-labeled data. Because reliable labels are expensive and community labels are noisy and survey-dependent, the authors train on injected transients combined with bogus-dominated survey data, remaining robust under strong label contamination. The method uses two networks trained with asymmetric co-teaching to handle categories with different label-noise levels, and analyzes learned representations through latent-space visualization. A low-cost hybrid uncertainty quantification (UQ) strategy leverages the dual-network setup to improve calibration, delivering calibrated uncertainty estimates. The authors, including Raphaël Bonnet-Guerrini, Bruno Sanchez, and Dominique Fouchez, show that injection-driven weak supervision enables scalable and consistent real-bogus classification without manual labels.

Paper Overview

Field: AI / Astronomy Authors: Raphaël Bonnet-Guerrini, Bruno Sanchez, Dominique Fouchez, Benjamin Racine, Maya Guy, Mariam Sabalbal, Manal Yassine, Vincenzo Piuri Published: 2026-07-06 arXiv: 2607.05393

Summary

Time-domain surveys produce large numbers of transient source candidates, and real-bogus classification is a critical step in automated discovery pipelines. Reliable labels are costly to obtain, while community labels tend to be noisy and survey-dependent.

This paper develops a real-bogus classification framework that requires no human-labeled data. It is trained using injected transients together with bogus-dominated survey data, remaining robust under strong class contamination while providing calibrated uncertainty quantification.

Key techniques

Asymmetric co-teaching: a dual-network model handles categories with different label-noise levels.
Latent-space visualization: used to analyze and interpret the learned representations.
Hybrid uncertainty quantification (UQ): a low-cost strategy that leverages the dual-network setup to improve calibration.

Conclusion

The results demonstrate that injection-driven weakly supervised training enables scalable and consistent real-bogus classification without any human labels.

---

*Auto-collected on 2026-07-06.*

Interpretable, Human-Label-Free Deep Learning for Real-Bogus Classification in Time-Domain Surveys

Summary

Paper Overview

Summary

Key techniques

Conclusion

Tags